[llvm-dev] Load combine pass

Sanjoy Das via llvm-dev llvm-dev at lists.llvm.org
Wed Sep 28 17:23:32 PDT 2016

Hi Artur,

Artur Pilipenko via llvm-dev wrote:
 > One of the arguments for doing this earlier is inline cost
 > perception of the original pattern. Reading i32/i64 by bytes look much
 > more expensive than it is and can prevent inlining of interesting
 > function.

I don't think this is just a perception issue -- if the loads have not
been widened then inlining the containing function _is_ expensive, and
the inliner cost analysis is doing the right thing.

 > Inhibiting other optimizations concern can be addressed by careful
 > selection of the pattern we’d like to match. I limit the
 > transformation to the case when all the individual have no uses other
 > than forming a wider load. In this case it’s less likely to loose
 > information during this transformation.

I agree -- I think widening loads in "obvious" cases like:

   i16 *a = ...
   i32 val = a[i] | (a[i + 1] << 16)

is more defensible than trying to widen the example that broke in

Regarding atomicity, the only real optimization that we'll lose (that
I can think of) is PRE.  Additionally, it may be more expensive to
lower wider atomic loads / stores, but that can be indicated by a
target hook.  For instance, on x86, I don't think:

   load atomic i16, i16* %ptr, unordered

is cheaper than

   load atomic i32, i32* %ptr.bitcast, unordered

so from a lowering perspective there is no reason to prefer the former.

Given this, perhaps scheduling load widening after one pass of GVN/PRE
is fine?

-- Sanjoy

 > I didn’t think of atomicity aspect though.
 > Artur
 >> On 28 Sep 2016, at 18:50, Philip Reames<listmail at philipreames.com> 
 >> There's a bit of additional context worth adding here...
 >> Up until very recently, we had a form of widening implemented in 
GVN.  We decided to remove this in https://reviews.llvm.org/D24096 
precisely because its placement in the pass pipeline was inhibiting 
other optimizations. There's also a major problem with doing widening at 
the IR level which is that widening a pair of atomic loads into a single 
wider atomic load can not be undone. This creates a major pass ordering 
problem of its own.
 >> At this point, my general view is that widening transformations of 
any kind should be done very late.  Ideally, this is something the 
backend would do, but doing it as a CGP like fixup pass over the IR is 
also reasonable.
 >> With that in mind, I feel both the current placement of LoadCombine 
(within the inliner iteration) and the proposed InstCombine rule are 
 >> Philip
 >> On 09/28/2016 08:22 AM, Artur Pilipenko wrote:
 >>> Hi,
 >>> I'm trying to optimize a pattern like this into a single i16 load:
 >>>   %1 = bitcast i16* %pData to i8*
 >>>   %2 = load i8, i8* %1, align 1
 >>>   %3 = zext i8 %2 to i16
 >>>   %4 = shl nuw i16 %3, 8
 >>>   %5 = getelementptr inbounds i8, i8* %1, i16 1
 >>>   %6 = load i8, i8* %5, align 1
 >>>   %7 = zext i8 %6 to i16
 >>>   %8 = shl nuw nsw i16 %7, 0
 >>>   %9 = or i16 %8, %4
 >>> I came across load combine pass which is motivated by virtualliy 
the same pattern. Load combine optimizes the pattern by combining 
adjacent loads into one load and lets the rest of the optimizer cleanup 
the rest. From what I see on the initial review for load combine 
(https://reviews.llvm.org/D3580) it was not enabled by default because 
it caused some performance regressions. It's not very surprising, I see 
how this type of widening can obscure some facts for the rest of the 
 >>> I can't find any backstory for this pass, why was it chosen to 
optimize the pattern in question in this way? What is the current status 
of this pass?
 >>> I have an alternative implementation for it locally. I implemented 
an instcombine rule similar to recognise bswap/bitreverse idiom. It 
relies on collectBitParts (Local.cpp) to determine the origin of the 
bits in a given or value. If all the bits are happen to be loaded from 
adjacent locations it replaces the or with a single load or a load plus 
 >>> If the alternative approach sounds reasonable I'll post my patches 
for review.
 >>> Artur
 > _______________________________________________
 > LLVM Developers mailing list
 > llvm-dev at lists.llvm.org
 > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

More information about the llvm-dev mailing list